Re: BBU Cache vs. spindles

On Fri, 29 Oct 2010, James Mansion wrote:

> Tom Lane wrote:
>> Uh, no, it is not. The difference is that we can update a byte in a
>> shared buffer, and know that it *isn't* getting written out before we
>>
> Well, I don't know where yu got the idea I was refering to that sort of thing
> - its
> the same as writing to a buffer before copying to the mmap'd area.
>> It's true that we don't know whether write() causes an immediate or
>> delayed disk write, but we generally don't care that much. What we do
>>
> Which is what I was refering to.
>> care about is being able to ensure that a WAL write happens before the
>> data write, and with mmap we don't have control over that.
>>
>>
> I think you have just the same control either way, because you can only force
> ordering
> with an appropriate explicit sync, and in the absence of such a sync all bets
> are off for
> whether/when each disk page is written out, and if you can't ensure that the
> controller
> and disk are write through you'd better do a hardware cache flush.too, right?
>
> A shame that so many systems have relatively poor handling of that hardware
> flush.

the issue is that when you update a mmaped chunk of data, it could be
written out immediatly without you doing _anything_ (and thanks to
multiple cores/threads, it could get written out while you are still in
the middle of updating it). When you update an internal buffer and then
write that, you know that nothing will hit the disk before you issue the
write command.

David Lang